metadata

license: mit
datasets:
  - name: cloud_sky_vis
    task:
      type: image-classification
    files:
      - name: image_metadata.csv
        format: csv
        split: train
        features:
          - name: index
            dtype: int32
          - name: filename
            dtype: string
          - name: label
            dtype: string
          - name: transform_shape
            dtype: string
          - name: transform_min
            dtype: float32
          - name: transform_max
            dtype: float32

Cloud and Sky Image Tensors for Classification

This dataset is designed for those interested in cloud classification projects. Due to licensing restrictions, the raw images cannot be shared publicly. However, the transformed tensors provided here are optimized for image classification tasks and are typically all you need for such projects.

Tensor Specifications

These tensors are preprocessed and normalized for use with ResNet models. The normalization parameters are as follows:

Mean: [0.485, 0.456, 0.406]
Standard Deviation: [0.229, 0.224, 0.225]

Cloud Class Labels

The dataset uses standard World Meteorological Organization (WMO) cloud classification labels, with the addition of "Clr" for clear skies. The labels used are:

As: Altostratus
Cb: Cumulonimbus
Cc: Cirrocumulus
Ci: Cirrus
Cs: Cirrostratus
Ct: Contrails
Cu: Cumulus
Ns: Nimbostratus
Sc: Stratocumulus
St: Stratus
Ac: Altocumulus
Clr: Clear Sky

For more information on cloud classification, please refer to the WMO Cloud Atlas: WMO Cloud Classification.

Dataset Structure

train_images_tensor.pt: Tensor of transformed images.
train_labels_tensor.pt: Corresponding labels for the images.
image_metadata.csv: Metadata detailing each image's properties.

Usage

These tensors are ready for direct use in training or testing ResNet and similar models for cloud classification tasks. I hope this dataset supports your research and projects.

To load the dataset:

import torch
import requests
from io import BytesIO

base_url = "https://huggingface.co/datasets/jcamier/cloud_sky_vis/resolve/main/"
tensor_url = f"{base_url}train_images_tensor.pt"
label_url = f"{base_url}train_labels_tensor.pt"

# Load the tensors
tensors = torch.load(BytesIO(requests.get(tensor_url).content))
labels = torch.load(BytesIO(requests.get(label_url).content))

# Verify the data is loaded correctly
print(tensors.shape)
print(labels.shape)

Enjoy!